after some testing
I found this and tested it (GGUF I-Matrix). Initially I was sceptical, but I was instantly surprised by the generated output. Even after long conversations (context size 24,576), it was quite good at bringing up stuff from the past and complaining that something had been mentioned incorrectly. Regarding the unscrupulous part, I agree... if you ask about certain things, it can be quite disturbing... my Character wanted to crush my head with a boot and cut me open... lol.
But I also have to say that after a longer conversation, the {char} started acting like it really felt something for me and refused to be aggressive... that was a very nice and surprising reaction from an LLM that wanted to kill me at the beginning. I also have the impression that the LLM did not drive the whole conversation towards sex, as many other models do.
Overall, I think you did a great job with the merge and the choice of models! For a first merge, I'm impressed!
P.S. To add something negative, I had some issues with the formatting, but they were very minor and cosmetic!
P.P.S. Talking in my name was rare! I edited the LLM message, removing what it said for me, and wrote in my reply, '@LLM Do not speak or act for me ever'. In my opinion, this helped; out of 100 messages, I only had the issue maybe five times at most.
I found this and tested it (GGUF I-Matrix). Initially I was sceptical, but I was instantly surprised by the generated output. Even after long conversations (context size 24,576), it was quite good at bringing up stuff from the past and complaining that something had been mentioned incorrectly. Regarding the unscrupulous part, I agree... if you ask about certain things, it can be quite disturbing... my Character wanted to crush my head with a boot and cut me open... lol.
But I also have to say that after a longer conversation, the {char} started acting like it really felt something for me and refused to be aggressive... that was a very nice and surprising reaction from an LLM that wanted to kill me at the beginning. I also have the impression that the LLM did not drive the whole conversation towards sex, as many other models do.
Overall, I think you did a great job with the merge and the choice of models! For a first merge, I'm impressed!P.S. To add something negative, I had some issues with the formatting, but they were very minor and cosmetic!
P.P.S. Talking in my name was rare! I edited the LLM message, removing what it said for me, and wrote in my reply, '@LLM Do not speak or act for me ever'. In my opinion, this helped; out of 100 messages, I only had the issue maybe five times at most.
Care to share your settings? i thought using my usual preset would work but im struggling to find settings that work for me.
I am no expert in those setting things ;) and I do not use Silly taver, if that is what you mean ;)
So i can tell you i use a temp of 0.7 -0.9 and the rest is unchanged to default, like rep penalty 1.05 / min-P 0.1
Sorry if i can not provide more help in that ;)